Attraction-Area based Geo-Clustering for LTE VehicularCrowdSensing Data Offloading

Douglas F. S. NunesUniversity of Sao Paulo, Brazil

douglas@usp.br

Edson S. MoreiraUniversity of Sao Paulo, Brazil

edson@icmc.usp.br

Bruno Y. L. KimuraFederal University of Sao Paulo, Brazil

bruno.kimura@unifesp.br

Nishanth SastryKing’s College London, UKnishanth.sastry@kcl.ac.uk

Toktam MahmoodiKing’s College London, UKtoktam.mahmoodi@kcl.ac.uk

ABSTRACTVehicular CrowdSensing (VCS) is one of the most emerging andpromising solutions designed to remotely collect data from smartvehicles. It enables a dynamic and large-scale phenomena monitor-ing just by exploring the variety of technologies which have beenembedded in modern cars. However, VCS applications might gener-ate a huge amount of data traffic between vehicles and the remotemonitoring center, which tends to overload the LTE networks. Inbig cities, this issue can be even more evident, given the numberof vehicles that may roam around. In this paper, we describe andanalyze a gEo-clUstering approaCh for Lte vehIcular crowDsEnsingdAta offloadiNg (EUCLIDEAN). It takes advantage of opportunis-tic vehicle-to-vehicle (V2V) communications to support the VCSdata upload process, preserving, as much as possible, the cellularnetwork resources (i.e., channels and bandwidth). In general, it isshown from the presented results that our proposal is a feasible andan effective scheme to reduce up to 92.98 % of the global demand forLTE transmissions while performing vehicle-based sensing tasks inurban areas. The most encouraging results were perceived mainlyunder high-density conditions (i.e., above 125 vehicles/km2), whereour solution provides the best benefits in terms of cellular networkdata offloading.

KEYWORDSLTE Data Offloading, Vehicular CrowdSensing, Vehicle Ad hocNetworks, Vehicular Geo-Clustering, VANET

1 INTRODUCTIONThe variety of technologies which have been incorporated in mod-ern vehicles is bestowing them the ability to act as amazing MobileSensor Platforms (MSP) [18]. Apart from having an increasing num-ber of sensors, these devices are also being equipped with a greatercomputing power and different wireless communication interfaces.

Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. Copyrights for components of this work owned by others than ACMmust be honored. Abstracting with credit is permitted. To copy otherwise, or republish,to post on servers or to redistribute to lists, requires prior specific permission and/or afee. Request permissions from permissions@acm.org.MSWiM’17, Nov 2017, Miami, USA© 2017 Association for Computing Machinery.ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00https://doi.org/10.1145/nnnnnnn.nnnnnnn

With such features, this generation of connected vehicles is repre-senting a wide dynamic sensing opportunity to build innovativeIntelligent Transportation Systems (ITSs) and a fruitful terrain fordeveloping Vehicular CrowdSensing (VCS) solutions towards theSmart City (SC) [12, 17] era. VCS [14, 18] is one of the most emerg-ing and promising schemes designed to remotely collect data (e.g.,fuel consumption, GPS position, engine status and speed) fromsmart vehicles. It allows for the deployment of powerful monitor-ing systems just by exploring the technologies embedded in MSPs,with no need for installing extensive infrastructure [5].

Using their native wireless communication capabilities, vehiclescan make their on-board data available directly to other local ve-hicles through Vehicle-to-Vehicle (V2V) ad hoc transmissions, orto the Internet via Vehicle-to-Infrastructure (V2I) transmissions.Such an infrastructure might be twofold: (i) computational stationsplaced alongside the roads, namely Road Side Units (RSUs), whichare IEEE 802.11p [10] access points envisioned to assist vehicularcommunication processes; and (ii) LTE radio base stations fromcellular networks deployed around. It is important to highlight herethat in the vehicular environment, the short-range wireless com-munications are provisioned by a set of standards stacked into aspecific architecture known as IEEE WAVE [20]. WAVE was createdto deal with the harsh communication conditions inherent in thisscenario (e.g., short contact time, high node speed, and high nodemobility), where the traditional Wi-Fi provides a poor experience.When employing local ad hoc transmissions to exchange data, thesmart vehicles get into a specific network formation called Vehic-ular Ad hoc Network (VANET) [19, 21]. Such kind of network isconsidered a key component to support the building of ITSs andVCS’s systems.

Under high-density condition in big cities, the VCS applicationsmight generate a huge amount of traffic between vehicles and mon-itoring center, tending to overload the network. Once the dataupload is usually accomplished via LTE, massive amounts of trans-missions can considerably degrade the Quality of Services (QoS)it offers [4]. To deal with this issue, we drew a gEo-clUsteringapproaCh for Lte vehIcular crowDsEnsing dAta offloadiNg (EU-CLIDEAN). EUCLIDEAN is focused on exploring opportunistic V2Vcommunications as a strategy to relieve LTE uplink during VCSdata acquisition. Firstly, stationary geo-clusters are created insidethe target area. Next, a subset of vehicles is selected to be in chargeto gather local information and to report them towards the moni-toring center. According to the simulation results of our proposal,in highly crowded conditions up to 92.98 % of vehicles were able

MSWiM’17, Nov 2017, Miami, USA

to report their VCS data samples without the need to access theLTE infrastructure. The two major contributions of this paper areas follows: (a) to evaluate the potential reduction of transmissionsover cellular network when V2V is additionally used to supportthe data upload process; and (b) to provide a vehicle geo-clusteringformation strategy in order to decrease the overall demand for LTEresources in VCS.

The remainder of this paper is organized as follows: Section 2reviews the main related works; Section 3 describes our proposal;Section 4 goes into finer details with respect to performed simu-lations; the experimental results are discussed in Section 5; and,finally, Section 6 provides our conclusions and future works.

2 RELATEDWORKSVCS is a compelling mean to deploy monitoring applications inhighly dynamic scenarios. This explains why this paradigm hasgained evidence in recent scientific research involving vehicularsensing. In line with our proposal, we segmented some of the ITSand VCS related works in two main groups.

The first one is represented by investigations focused on clus-tering vehicles in a centralized or decentralized fashion [2, 3, 15].According to [8, 13] cluster-based networking can be consideredan attractive solution to provide spatial reuse of the bandwidth,to reduce network congestion, and to simplify message routing ordata fusion methods. The cluster formation process reported by theworks belonging to this group usually involves several steps andtransmissions of additional packets to (i) ensure that the createdcluster is the best formation; to (ii) ensure that the current clusterformation is consistent and updated, dealing with the vehicles’ mo-bility; and to (iii) guarantee the election and eventual changes ofcluster heads. These cluster formation and maintenance messagesmay incur in an extra and considerable overhead to the VANET,since the vehicles have a high dynamic mobility pattern and theVCS natively requires multiple ad hoc transmissions to exchangetheir sensing data itself. To avoid those additional packets, we de-signed a new geographic clustering approach where the vehiclesare able to create clusters and to select their respective cluster headsin a straightforward manner.

The second group brings together studies on provisioning LTEoffloading techniques to drain VCS data over alternative networks[4, 7, 15, 16]. The performance results shown by these works demon-strate that the transmissions over cellular radio access system canbe drastically reduced when VANET and LTE technologies arecombined to support VCS tasks. Our proposal is aligned with theabove-mentioned researching, insofar as it integrates some featuresof those approaches while it extends and creates others for address-ing LTE network overload issues. In the majority of the workscentered on cellular network offloading, RSUs are fixed on the en-vironment in order to directly collect sensing information fromvehicles or/and to forward them to the Internet via a high-speedlink. It is important to highlight, however, that this vehicular com-munication infrastructure is expensive. Deploying them in largescale urban areas, e.g., can be infeasible due to its high cost. We arenot considering RSUs in our opportunistic LTE offloading scheme.To relieve the cellular network in crowded conditions, with a highvolume of VCS data implied, we explore V2V communications to

gather in situ sensing information and to reduce the number of LTEtransmissions needed for uploading those data to the monitoringcenter.

3 EUCLIDEANThe following definitions are used throughout this Section to sup-port the explanation of our proposal:

Definition 1 (Region of Interest). Region of Interest (ROI) is a well-defined geographical area (e.g., a portion of a city) whose correlatedinformation is relevant to the sensing applications.

Definition 2 (Attraction Area). In this paper, Attraction Area is alogical, circular, and stationary geographic region within an ROIwhose centroid is the point where it is expected to find an LTErelay.

Definition 3 (Geo-Cluster). A Geo-Cluster is a grouping of vehi-cles created within an Attraction Area so that they can cooperateamongst one another and exchange their data.

For an illustrative setting, let’s consider an urban-like area anda VCS system which is focused on acquiring data about the roadtraffic itself, via en route vehicles monitoring. Traffic managementauthorities, drivers, autonomous vehicles or even passengers arethe potential stakeholders. In this scenario, we assume that allvehicles are equipped with an On-Board Unit (OBU), a Global Posi-tioning System (GPS) receiver, an LTE cellular network interface,an IEEE 802.11p short-range wireless communication interface, andsome built-in machine information sensors. We also assume thatthe VCS system encompasses a client application version, namedas VCS_TClientApp, which is deployed on each OBU, and a corre-sponding server application version, named as VCS_TServerApp,which is hosted on a cloud server environment, here called moni-toring center.

Before starting a data collection, the acquiring task propertiesneed to be configured by the VCS_TServerApp. This process con-sists of assigning values to parameters, such as data sample collectionfrequency, data report frequency, sensing area, kind of data required,and target vehicles. According to the VCS application domain, theseparameters can assume distinct values. In road traffic monitoring,for instance, a higher data sample collection frequency and a higherdata report frequency are expected, in order to be able to catch therecent road traffic changes. Sensing weather conditions in a largearea, on the other hand, may be less strict and may require a lowerfrequency value for these parameters.

After the acquisition task has its properties defined, an instanceof it will be sent out to each desired vehicle roaming in the ROI. Oncethe task has been received by a target vehicle, the VCS_TClientAppwill collect the required on-board data and then, under certain pre-defined conditions (e.g., a time window or a buffer threshold), itwill report (upload) them to the VCS_TServerApp. This upload op-eration is depicted in the Fig. 1a, where the data are acquired usingonly V2I LTE transmissions. This is the main approach adopted bythe majority of applications which employ the VCS [16].

When a large number of vehicles is considered to perform VCStasks (e.g., in big cities), a huge amount of data is expected to requireLTE uplink resources in order to traverse the network towards the

Attraction-Area based Geo-Clustering for LTE Vehicular CrowdSensing Data Offloading MSWiM’17, Nov 2017, Miami, USA

(a) data upload using only V2I LTE (b) data upload using V2V and V2I LTE

Figure 1: VCS data uploading: (a) using only V2I LTE; and (b) using V2V and V2I LTE

monitoring center. Since those LTE resources are finite, this condi-tion tends to significantly degrade the QoS that this infrastructureoffers. Under exceptional circumstances, the LTE network mighteven break down. This is not an unrealistic assumption if we takeinto account that we are just entering the Internet of Things (IoT)[22] age and that a plenty of devices, besides vehicles, will also com-pete for LTE resources in the near future. It means that the paradigmillustrated by Fig. 1a does not scale well in very crowded environ-ments. Therefore, in an attempt to reduce this negative impact, weare proposing a strategy to save LTE uplink resources by means ofthe use of opportunistic V2V communications in addition to the tra-ditional approach presented by Fig. 1a. A pictorial representation ofour scheme can be seen in Fig. 1b. Instead of each vehicle uploadingits own data directly to the VCS_TServerApp, they will send themto a local subset of vehicles which will be in charge of doing that.The proposed approach, named as EUCLIDEAN, operates over twonovel algorithms particularly designed for this purpose: (i) Station-ary Attraction Areas Allocation’s Algorithm; and (ii) AttractionArea-based Clustering Protocol.

3.1 Stationary Attraction Areas AllocationAlgorithm (S4A)

Finding a subset of vehicles that will be considered LTE relays inVCS systems is not a trivial procedure. Once the sensing is usuallyperformed inside an ROI, defining which vehicles will be selectedas LTE relays involves considering where they are located as well.Meeting these two requirements is a challenge, mainly because theVANETs topologies change all the time. Moreover, spreading relaysuniformly within the ROI is relevant to guarantee that all vehicleswill be able to take advantage of the use of V2V communicationsto upload their VCS data, thus increasing the potential of savingLTE uplink resources. In this sense, instead of designing a strat-egy centered on discovering the best vehicle to act as a relay, ourscheme is focused on determining where it is expected to find arepresentative one to drain such data over LTE network.

For this end, we are proposing an effective method based on theconcept of logical attraction area (See Definition 2). All vehiclesinside an attraction area will send their data to a correspondingLTE relay via V2V transmissions, and then it will report that in-formation to the VCS_TServerApp using V2I LTE transmissions, asrepresented by “Attraction area 1" in Fig. 2.

Figure 2: EUCLIDEAN illustrative representation.

Whereas a single LTE relay is expected for each attraction area,we defined a radius for it with the same value of the V2V wirelesstransmission technology range used by vehicles (i.e., accordingto the IEEE 802.11p interface properties). Thus, if an LTE relay islocated just above a given centroid, it will be able to communi-cate with all other vehicles inside that related attraction area. Thisassumption is relevant because the attraction area amplitude inEUCLIDEAN implies directly in the number of LTE relays selected.The smaller the attraction area, the greater the number of LTErelays which are expected inside the ROI. However, by reducingthe number of LTE relays, we are also reducing the number of LTEtransmissions so the benefits of this approach tend to increase. The

MSWiM’17, Nov 2017, Miami, USA

LTE relay selection algorithm will be better detailed in the nextsubsection.

To geo-locate the attraction areas, the S4A first maps all the co-ordinates of the boundaries of the ROI. Next, it starts to place theircentroids side-by-side along the x andy-axis, until the whole ROI iscovered. The attraction areas may be totally disjoint among them orthey can be distributed in such a way that a given portion of themare overlapped, e.g., they are slightly overlapped in Fig. 2. Therefore,we must provide the allocation algorithm with the bordering coor-dinates of the ROI and the desired distance amongst the centroids.This can be done by using an interface of the VCS_TServerAppwitha digital map of the city, by means of clicking and dragging overthe area of interest, or via text fields. The S4A was designed toautomate the process of determining how much attraction areaswill be needed and where they should be placed. It is carried out onthe server side, i.e., by the VCS_TServerApp, before spreading outthe VCS tasks. Performing the S4A on the server side allows theVCS_TServerApp to rearrange all the attraction areas at any time.Thus, the vehicle-based sensing becomes even more versatile. Afterthis step, a list of the attraction area centroid coordinates as wellas the assumed radius will be sent to the VCS_TClientApps alongwith the VCS tasks themselves. Each vehicle will keep these datastored in its local database, for they are imperative to support thegeo-clustering formation. A pseudo code description of the S4A isshown in Algorithm 1.

Algorithm 1: S4AInput: A set ROIbc = {bc1(x,y) ,bc2(x,y) , . . . ,bcn(x,y) } of

coordinates of the boundaries of the ROI; and apositive distance value between the centroids.

Output: A set AAcc = {cc1(x,y) , cc2(x,y) , . . . , ccn(x,y) } ofattraction areas centroids coordinates.

1 Coordy ← TopMaxCoordy (ROIbc );2 i ← 0;3 while Coordy ≥ BottomMinCoordy (ROIbc ) do4 Coordx ← Le f tMinCoordx (ROIbc );5 while Coordx ≤ RiдhtMaxCoordx (ROIbc ) do6 cci(x,y) ← (Coordx ,Coordy );7 appendToAAcc (cci(x,y) );8 i ← i + 1;9 Coordx ← Coordx + distance;

10 Coordy ← Coordy + distance;11 return AAcc ;

3.2 Attraction Area-based Clustering Protocol(AACP)

As introduced in Section 2, cluster-based networking is an appeal-ing solution to provide spatial reuse of the bandwidth, to reducenetwork congestion, and to simplify the routing of packets or theaggregation/fusion of data. Considering that these advantages arerelevant to the VCS domain, we considered building ephemeralgeo-clusters (See Definition 2) of vehicles in our approach. Withthis grouping formation, the VCS applications can decrease the

number of LTE data transmissions needed to perform their sensingtasks. In the AACP, each geo-cluster has associated with it a setof vehicles named Cluster Members (CMs) and a representativeone called Cluster Head (CH). The CH is a vehicle timely selectedto be responsible for reporting VCS data to the monitoring centeron behalf of its one-hop neighbors. In this sense, only CHs areexpected to use LTE resources and not all vehicles inside the ROI.Hereafter, the term CH will be used to refer to LTE relay mentionedin the last subsection.

In order to create a geo-cluster, the vehicles make use of: (i)their GPS coordinates; (ii) the GPS coordinates of their neighbors;(iii) the coordinates of the centroids; and (iv) the radius of theattraction areas. The own GPS coordinates are obtained from the on-board GPS receiver, the neighbors’ GPS coordinates are known viabeaconing services1 used by vehicles, and the last two parametersare provided by the VCS_TServerApp, as previously mentioned.

Each vehicle vi ,∀v ∈ V , keeps a local database with the listof centroids mapped to that ROI, denoted by C, and an updatedtable with the ID and the current GPS coordinates of all its neigh-bors, denoted by N. With those data, a vehicle is able to find itsnearest centroid, set the geo-cluster it belongs to, and select itscorresponding CH in a distributed and straightforward way. Thenearest centroid is discovered by means of the Euclidean distancebetween the vehicle and centroid,

d(vi , c j ) =√(vix − c jx )2 + (viy − c jy )2, (1)

where, {vix ,viy } and {c jx , c jy } are the GPS coordinates mappedinto the Cartesian coordinates of the vehiclev and the centroid c , re-spectively. The nearest centroid to the vehiclevi will be that c j ,∀c ∈C, with the shortest distance dmin = min{d1(vi , c j ), . . . ,dn (vi , c j )}.As the AACP assumes that all vehicles roaming up to radius metersfar from a centroid will belong to the same geo-cluster, by knowingthe nearest c j as well as the current position of its neighbors N, avehiclevi becomes conscious of its geo-cluster formation. However,a CH still needs to be defined. A vehicle vi will be considered aCH of its geo-cluster if it has the shortest distance dmin, among itsneighbors N, to the corresponding centroid c j (See Attraction area2 in Fig. 2).

When knowing the coordinates and IDs of its neighbors, a vehi-cle can calculate by itself the distances of each CM to the nearestcentroid. Thus, all vehicles belonging to a geo-cluster are able toidentify their respective CH in a distributed way, at any time, with-out the need to exchange additional messages for this. If CHs moveaway from their centroids, due to their mobility, new CHs might beselected, once all vehicles inside a geo-cluster are able to becomeCHs. In this context, EUCLIDEAN considers that centroids are, actu-ally, attraction points towards which surrounding VCS data shouldflow, in the attempt to achieve access to the LTE infrastructure. Apseudo code description of the AACP is shown in Algorithm 2.

During the data collection process, every vehicle periodicallyperforms the following steps:(s1) Read the required on-board data samples and insert them into

its local VCS_message_buffer.1We assume that all OBUs periodically broadcast their identity (ID) and GPS locationin special packets denoted beacons, with the aim of making the vehicles aware of allits neighbors in quasi real time.

Attraction-Area based Geo-Clustering for LTE Vehicular CrowdSensing Data Offloading MSWiM’17, Nov 2017, Miami, USA

(s2) Check if it is the current CH. If so, it will send the data sam-ples stored in the VCS_message_buffer directly to the remoteVCS_TServerApp. Otherwise, it will send all data samples to itsCH. After receiving an ack message from the CH, the vehiclewill discard from its buffer those samples that were uploadedvia V2V. This latter ad hoc data exchange process employs onlysingle-hop transmissions. To avoid a continuous flip-flop datasamples exchange between CHs inside a geo-cluster, when anew one is selected, a maximum buffer length was defined. Ev-ery time this threshold is reached, the CH will relief its buffersending all data samples to the remote monitoring center viaV2I LTE.

(s3) Start a timer to the next reading. Once the timer expires (time-out), go back to the step (s1).

Algorithm 2: AACPInput: The vehicle GPS coordinates vi(x,y) ; a set

N = {n1(x,y) ,n2(x,y) , . . . ,nn(x,y) } of GPS coordinatesfrom its neighbors; a setC = {c1(x,y) , c2(x,y) , . . . , cn(x,y) } of centroidscoordinates; and a positive radius value.

Output: A geo-cluster affiliation ID (GCAID ); and the clusterhead ID (CHID ).

1 nearestCentroid ← c1(x,y) ;2 forall c j ∈ C do3 d ← eucliDistance(vi , c j );4 if d < eucliDistance(vi ,nearestCentroid) then5 nearestCentroid ← c j(x,y) ;

6 if eucliDistance(vi ,nearestCentroid) < radius then7 GCAID ← nearestCentroidID ;8 CHID ← vi ;9 dmin ← eucliDistance(vi ,nearestCentroid);

10 forall ni ∈ N do11 d ← eucliDistance(ni ,nearestCentroid);12 if d < dmin then13 dmin ← d ;14 CHID ← ni ;

15 return GCAID ,CHID ;

Vehicles that were nominated as CH can also aggregate andconvert the gathered data into an average value before reporting itto the VCS_TServerApp. In this sense, the volume of data uploadedover LTE can be drastically reduced. On the other hand, even thatVCS_TServerApp asks for a genuine data sample of each vehicleinvolved with VCS, the advantage of using a single header stack fora set of gathered data is not negligible, as will be shown in Section5. We adopted this latter model in our experiments. The discussionabout the data fusion technique is out of the scope of this work.

4 SIMULATION DESIGN4.1 Simulation SetupIn order to evaluate the performance of EUCLIDEAN, we createdsome experiments using a suite of state-of-art simulation tools.

The first one was the OMNeT++ [1], an extensible, modular, andcomponent-based C++ framework to support network simulations.The SUMO [11] was employed to model our urban-like scenarioand to model our vehicular transport system, i.e., defining vehicles,roads, traffic lights, and the vehicular traffic itself. Lastly, we madeuse of VEINS LTE [9] to couple the network models provided byOMNET++ with the vehicular traffic models provided by SUMO.VEINS LTE is an open source framework that was created based onOMNET++ and SUMO to carry out vehicular network simulations. Asummary of the main network and scenario simulation inputs usedby our experiments are shown in Table 1 and Table 2, respectively.

Table 1: Network Simulation Parameters and Values

Parameter ValueV2V Packet format WAVE short messageV2V beaconing Frequency 1HzV2V transmission range 500mVANET MAC, PHY technology IEEE 802.11p802.11p Carrier frequency 5.9 GHz802.11p Transmission Power 20mW802.11p Sensitivity -89 dBm802.11p MAC bit rate 18Mbps802.11p Radio Propagation Model Simple path loss modelVCS data report Interval 30 sSimulation duration time 300 sLTE transport layer Protocol UDPLTE MAC Queue size 2MBLTE Carrier frequency 2.1 GHz

So that we could reproduce a realistic urban-like scenario, wedecided to use a ≈ 4 km2 portion of the Bologna city [6] (See Fig. 3).In our experiments, the densities of the vehicles change over timeassuming values from 1 to 500; also, the maximum speed reachedby them was 70 km/h. The cellular network services are provided byone LTE micro-cell antenna (eNobeB), placed in a central positionof the scenario, in such a way it can cover the entire area.

Table 2: Scenario Simulation Parameters and Values

Parameter ValueScenario area ≈ 4 km2 (1990m × 2150m)Vehicles densities [1, 500] vehiclesMaximum vehicles’ speed 70 km/h

4.2 Performance MetricsTo assist our analysis as well as to demonstrate the potential of ourproposal to reduce VCS data transmission over LTE network, wedefined the following performance metrics:

MSWiM’17, Nov 2017, Miami, USA

• Total LTE Packets Transmitted (TLPT): it is measuredby the sum of all VCS data packets transmitted over the LTEnetwork throughout the data acquisition process;• Samples per LTE Transmission (SLT): this metric is cal-culated by using the ratio between the number of VCS datasamples received by the VCS_TServerApp and the numberof LTE transmissions used to report those data;• Average Upload Delay (AUD): it is measured by the re-mote VCS_TServerApp, computing the average time elapsedbetween the instant of time when the VCS data were sam-pled inside the vehicles and the instant of time when theyreached the VCS_TServerApp; and• Number of Cluster Heads Selected (NCHS): this metricrepresents the total of CHs that were selected during all thesimulation time.

Figure 3: Simulation scenario: area in the city of Bologna [6]

The simulations were performed using multiple runs, with dif-ferent seed values. We plotted average results whose confidenceintervals were calculated with the significance level of 0.05, i.e.,assuming 95 % of confidence level.

4.3 Experimental MethodologyFor our experiments, it was considered the illustrative scenariopreviously presented by Section 3. We implemented an instanceof the VCS_TServerApp which spreads VCS tasks to acquire andreport the current GPS coordinates (vi(x,y) ) and the current speedof the vehicles (vi(speed ) ) every 30 s (this means that both datasample collection frequency and data report frequency were set to30 s). The whole Bologna’s scenario was assumed to be the ROI andall vehicles were considered by the VCS. To assist the geo-clusterformation, all vehicles periodically (by 1Hz 2) broadcast a HELLOpacket, via beaconing service, containing its ID and its currentgeographical coordinates. Different vehicle densities (1, 10, 50, 100,300, 500) 3 and three distinct attraction area centroids’ distances(500, 700, 1000) were taken into account by simulations: 500m forheavily overlapped attraction areas; 700m for slightly overlappedones; and 1000m for totally disjoint ones. The labels “V2V andV2I-500", “V2V and V2I-700", and “V2V and V2I-1000" are used in2Other frequency could be used. Evaluating the best one is out of the scope of thiswork.3This variation aims at representing those situations where the transit is fluent, withlow vehicle traffic per km2 , as well as those situations where traffic congestion mightbe happening. In other words, we aim at evaluating EUCLIDEAN in sparse and high-density vehicular flow conditions.

the next figures for referring to the distances of 500m, 700m, and1000m, respectively, among centroids. The performance of our LTEoffloading strategy was evaluated from individuals’ perspectives,mostly comparing it with the traditional pure V2I LTE approach.

5 PERFORMANCE ANALYSIS OFSIMULATION RESULTS

5.1 Traffic reduction from the LTE offloadingstrategy

Firstly, we were interested in knowing the potential reduction inthe number of transmissions over cellular network when V2V com-munications were also considered in the VCS data upload. Theresults presented in Fig. 4 show a substantial decreasing in higherdensity conditions and with slightly overlapped geo-clusters. Inthis setting, our strategy was able to spend only 7.02 % of the totalLTE transmissions used by the pure V2I LTE scheme to report alldata samples. Considering an IPv4 header size of 20 bytes (i.e., theshortest) and a UDP header size of 8 bytes, these savings in LTEtransmissions while using our geo-cluster approach represent upto 1.03Mbits4 of extra data traffic which were not sent over cellularnetworks. These benefits resulted from the fact that a single headerstack is used for a set of gathered VCS data. The aforementionedamount of extra data traffic is not negligible, mostly because itaccounts for almost two times the VCS data themselves obtainedduring the simulation duration time.

50 100 150 200 250 300 350 400 450 500

101

102

103

104

Number of Vehicles

TLPT

(inlogscale)

Pure V2I LTE

V2V and V2I-500

V2V and V2I-700

V2V and V2I-1000

Figure 4: TLPT vs. Number of Vehicles.

5.2 The Impact of Cluster AllocationWe were also interested in understanding how the variations inthe clusters allocation process would impact in the upload stage.According to the results depicted in Fig. 5, as expected, the ratioof samples sent to the VCS_TServerApp per V2I LTE transmissiontends to follow the increase of the vehicles’ densities. However, theratio is relatively reduced when the allocation of heavily overlap-ping geo-clusters is adopted. Since more CHs are present in thissetting, as shown in the Fig. 7, the volume of samples carried by4This means 4635 samples not sent directly via LTE × (20 bytes from IPV4 + 8 bytesfrom UDP).

Attraction-Area based Geo-Clustering for LTE Vehicular CrowdSensing Data Offloading MSWiM’17, Nov 2017, Miami, USA

each one is smaller. Besides that, the presence of more CHs in thescenario tends to increase the data upload frequency, reducing notonly the SLT, but the AUD as well (Fig. 5 and Fig. 6).

10 50 100 300 500

2

4

6

8

10

12

14

16

Number of Vehicles

SLT

Pure V2I LTE

V2V and V2I-500

V2V and V2I-700

V2V and V2I-1000

Figure 5: SLT vs. Number of Vehicles.

From the results of Fig. 5 and Fig. 6, we can observe a closerelation between the LTE packet payload length, translated into theSLT metric, and their average upload delays. The higher the SLT,the higher the AUD tends to be. The AUD in the geo-cluster ap-proach is also affected by the time the samples remain stored in theVCS_message_buffers. When a sample generated by CMs arrivesin the CH just after it reported its buffered messages, this samplewill be held by the CH until the next report event is triggered. Itis important to point out here that, according to the applicationrequirement, the delays can be reduced by means of mechanismswhich periodically check the oldest samples in the buffer. If it ex-ceeds a boundary, the CH may anticipate the data upload process.

10 50 100 300 5000

5

10

15

20

25

Number of Vehicles

AUD

(s)

Pure V2I LTE

V2V and V2I-500

V2V and V2I-700

V2V and V2I-1000

Figure 6: AUD vs. Number of Vehicles.

5.3 The Bunch of SamplesOur next analysis is focused on evaluating how the density ofvehicles and the geo-clusters allocation method (disjoint or over-lapped) would affect the manner the samples were transmitted in

LTE packets. Despite the assumption that the CHs are responsi-ble for gathering local samples and upload them inside a uniqueLTE packet, we noticed that in many LTE packets received by theVCS_TServerApp there was a single data sample. Thus, we makeuse of the term bunch of samples for referring to those LTE packetssent to the VCS_TServerApp, whose payloads contained more thana single data sample.

10 50 100 300 500

100

200

300

400

500

600

Number of VehiclesNum

bero

fCHs

V2V and V2I-500

V2V and V2I-700

V2V and V2I-1000

Figure 7: NCHS vs. Number of Vehicles.

The results plotted in Fig. 8 show an ascendant tendency tofind bunches of samples in scenarios with a higher density of ve-hicles. However, they were more present in heavily overlappedgeo-clusters. Similarly to the previous observation, this is related tothe number of CHs which exists in this condition (see Fig. 7). Com-paring these results with those depicted in Fig. 5, Fig. 6, and Fig. 7we can realize that to obtain smaller end-to-end delays in VCS, it isrelevant to consider using EUCLIDEAN with heavily overlappedgeo-clusters formation. On the other hand, if the intention is toreduce, as much as possible, the number of LTE transmission, weneed to increase the number of samples inside a LTE packet. Thisrequirement can be satisfied by using EUCLIDEAN with slightlyoverlapped geo-cluster formation or totally disjoint geo-clusterformation, since their results were somewhat similar in higherdensities conditions.

600800

1,000

200

4000

200

400

Centroids DistanceN. o

f Vehicl

es

Buncheso

fSam

ples

0

100

200

300

Figure 8: Bunch of samples transmitted over LTE uplink vs.Number of Vehicles.

MSWiM’17, Nov 2017, Miami, USA

5.4 VCS Samples Size and Storage OccupationIn the presented experiment, each vehicle was scheduled to generate2 data samples/min. Considering that a data sample is composed by thelatitude (4 bytes) and longitude (4 bytes) GPS coordinates values,plus the car speed (4 bytes) value, each vehicle was able to produce24 bytes/min (accounting for 12 bytes per data sample).

During our simulations, we noticed some CHs carrying up to130 data samples in their VCS_message_buffers. This was themaximum buffer length value reached during our observation;also, it represents a peak of ≈ 2KB 5 of memory occupation (avehicle ID (4 bytes) is associated with each data sample stored inVCS_message_buffer for retrieving the source of VCS data after-wards). Therefore, it is relevant to point out here that the distributednature of EUCLIDEAN preserves the storage capacity of the CHsby means of (i) scattering the VCS data samples among differentCHs and (ii) performing its periodic V2I LTE uploads.

6 CONCLUSIONThis paper describes and analyzes the EUCLIDEAN: a geo-clusteringapproach for LTE VCS data offloading. In this scheme, clusters ofvehicles are created within specific and stationary regions (attrac-tion areas) of the ROI. One CH is then selected for each geo-cluster.The VCS data are locally gathered by CHs and only this kind ofvehicle will be in charge to report them towards the monitoringcenter. Thus, exclusively a subset of vehicles can make use of V2ILTE transmissions. As a consequence, the traffic overhead getsshifted from the cellular network to the ad hoc inter-vehicles level.

According to our simulation results, up to 92.98 % of vehicleswere able to report their samples without transmitting them di-rectly to the monitoring center. The most encouraging resultswere perceived mainly under high-density conditions (i.e., above125 vehicles/km2. It means that exactly in those situations where theLTE network is prone to be overwhelmed with data transmissions,our proposal provides the best benefits in terms of cellular networkdata offloading. Future steps will consider deepening the researchin cellular offloading techniques. This includes: (i) improving EU-CLIDEAN so that it can be more dynamic, adapting the amplitudeof the attraction areas according to traffic flow changes; (ii) incor-porating data fusion methods to our approach; and (iii) exploitingoffloading in LTE downlink as well. Assuming that we are juststarting the IoT and SC era, innovative solutions to offload LTEnetworks, dealing with the massive amount of data generated andrequired by MSP, will be imperative in the next years.

REFERENCES[1] 2017. OMNeT++, a Discrete Event Simulator. (apr 2017). https://omnetpp.org/[2] Hamid Reza Arkian, Reza Ebrahimi Atani, Atefe Pourkhalili, and Saman Ka-

mali. 2014. Cluster-based traffic information generalization in Vehicular Ad-hocNetworks. Vehicular Communications 1, 4 (2014), 197 – 207.

[3] M. Azizian, S. Cherkaoui, and A. S. Hafid. 2016. A distributed D-hop clusterformation for VANET. In 2016 IEEE Wireless Communications and NetworkingConference. 1–6. https://doi.org/10.1109/WCNC.2016.7564925

[4] A. Bazzi, B. M. Masini, and G. Pasolini. 2012. V2V and V2R for cellular resourcessaving in vehicular applications. In 2012 IEEE Wireless Communications andNetworking Conference (WCNC). 3199–3203.

[5] Alessandro Bazzi and Alberto Zanella. 2016. Position Based Routing in CrowdSensing Vehicular Networks. Ad Hoc Netw. 36, P2 (Jan. 2016), 409–424.

5(12 bytes of each sample + 4 bytes of vehicle ID) × 130 buffer entries

[6] Laura Bieker, Daniel Krajzewicz, AntonioPio Morra, Carlo Michelacci, and FabioCartolano. 2015. Traffic Simulation for All: A Real World Traffic Scenario from theCity of Bologna. Springer International Publishing, Cham, 47–60.

[7] G. el Mouna Zhioua, J. Zhang, H. Labiod, N. Tabbane, and S. Tabbane. 2014.VOPP: A VANET offloading potential prediction model. In 2014 IEEE WirelessCommunications and Networking Conference (WCNC). 2408–2413. https://doi.org/10.1109/WCNC.2014.6952726

[8] Andrea Gorrieri, Marco MartalÚ, Stefano Busanelli, and Gianluigi Ferrari. 2016.Clustering and sensing with decentralized detection in vehicular ad hoc networks.Ad Hoc Networks 36, Part 2 (2016), 450 – 464.

[9] Florian Hagenauer, Falko Dressler, and Christoph Sommer. 2014. A Simulator forHeterogeneous Vehicular Networks. In 6th IEEE Vehicular Networking Conference(VNC 2014), Poster Session. IEEE, Paderborn, Germany, 185–186. https://doi.org/10.1109/VNC.2014.7013339

[10] D. Jiang and L. Delgrossi. 2008. IEEE 802.11p: Towards an International Stan-dard for Wireless Access in Vehicular Environments. In VTC Spring 2008 - IEEEVehicular Technology Conference. 2036–2040.

[11] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. 2012.Recent Development and Applications of SUMO - Simulation of Urban MObility.International Journal On Advances in Systems andMeasurements 5, 3&4 (December2012), 128–138.

[12] C. Kyriazopoulou. 2015. Smart city technologies and architectures: A literaturereview. In 2015 International Conference on Smart Cities and Green ICT Systems(SMARTGREENS). 1–12.

[13] C. R. Lin and M. Gerla. 1997. Adaptive clustering for mobile wireless networks.IEEE Journal on Selected Areas in Communications 15, 7 (Sep 1997), 1265–1275.

[14] Barbara Mavì Masini. 2016. Vehicular Networking for Mobile Crowd Sensing.Ad Hoc Netw. 36, P2 (Jan. 2016), 407–408.

[15] P. Salvo, I. Turcanu, F. Cuomo, A. Baiocchi, and I. Rubin. 2016. LTE floating cardata application off-loading via VANET driven clustering formation. In 2016 12thAnnual Conference on Wireless On-demand Network Systems and Services (WONS).1–8.

[16] R. Stanica, M. Fiore, and F. Malandrino. 2013. Offloading Floating Car Data. In 2013IEEE 14th International Symposium on "AWorld of Wireless, Mobile and MultimediaNetworks" (WoWMoM). 1–9. https://doi.org/10.1109/WoWMoM.2013.6583391

[17] K. Su, J. Li, and H. Fu. 2011. Smart city and the applications. In 2011 InternationalConference on Electronics, Communications and Control (ICECC). 1028–1031.

[18] Julian Timpner and Lars Wolf. 2016. Query-response Geocast for VehicularCrowd Sensing. Ad Hoc Netw. 36, P2 (Jan. 2016), 435–449.

[19] Y. Toor, P. Muhlethaler, A. Laouiti, and A. D. La Fortelle. 2008. Vehicle Ad Hocnetworks: applications and related technical issues. IEEE Communications SurveysTutorials 10, 3 (Third 2008), 74–88. https://doi.org/10.1109/COMST.2008.4625806

[20] R. A. Uzcategui, A. J. De Sucre, and G. Acosta-Marum. 2009. Wave: A tutorial.IEEE Communications Magazine 47, 5 (May 2009), 126–133.

[21] Yu Wang and Fan Li. 2009. Vehicular Ad Hoc Networks. Springer London, London,503–525.

[22] Feng Xia, Laurence T Yang, Lizhe Wang, and Alexey Vinel. 2012. Internet ofthings. International Journal of Communication Systems 25, 9 (2012), 1101.